build!: raise the standard to C++23, and spend it

[andiwand]

🤖 Generated with Claude Code

Raises the standard to C++23 and spends it on the three things this codebase
had waiting for it. Five commits, each reviewable on its own.

Rebased on main, now that #827 (the parts that needed no C++23) and #828 (the
locale fix, which brought fmt in) have landed. Everything left here genuinely
requires the standard.

The bump costs nothing

The bump commit is build configuration and documentation only — no source change
at all. The two transitive includes libc++'s C++23 headers stopped handing out
went in with #827.

odr + odr_test + translate, RelWithDebInfo	0 warnings
the same with -Werror, as build_test.yml runs it	clean
odr_test	1443 passed, 6 skipped, 0 failed — reference html byte for byte
clang-tidy on every changed file	no new finding; each one also fires on main, checked file by file
every changed source, compiled by NDK 28.1 to an object file	clean
odr built at CMAKE_OSX_DEPLOYMENT_TARGET=12.0	clean — fmt does not switch to std::to_chars at C++23, 0 undefined symbols
conan install -s compiler.cppstd=23	one dependency rebuilds (gtest); every other package ID is unchanged

MSVC is the caveat: it has no /std:c++23, and CMake maps CXX_STANDARD 23
to /std:c++latest. The windows-2022-msvc-1940 job moves onto C++26-draft mode.

Three ceilings sit under the standard

None is visible from a local build, so AGENTS.md records all three.

1. The library half stops at emsdk 3.1.73's libc++ 18.1 — the oldest in the
   profile matrix, and the newest emsdk conan-center packages, so not a pin that
   can be bumped. No views::enumerate, std::generator,
   std::move_only_function, std::flat_map. std::mdspan is out from the
   other side, since libstdc++ has none.
2. std::format is unusable, on every slice. The apple profiles deploy to
   macOS 12 / iOS 15 (apple.jinja), and libc++ annotates the floating-point
   std::to_chars that <format> formats through as introduced in macOS 13.3.
   <format> instantiates that formatter whatever the argument types are, so
   even std::format("{}:{}", host, port) fails — and one call anywhere in
   src/ breaks the framework build. A host-default compile passes and proves
   nothing; -mmacosx-version-min=12.0 is the check.
3. Deducing this is fine on accessors, not on a capturing recursive lambda.
   NDK 28.1's clang 19 segfaults on [&](this auto self, …) that recurses. Only
   codegen shows it — -fsyntax-only passes.

CI found 2 and 3 on the first push of this branch, after a local run that had
tested with -fsyntax-only and no deployment target. Both are now in AGENTS.md
with the command that reproduces them.

What it is spent on

Deducing this — 42 accessor pairs. #823 had just centralised the element
registry, and SideTable / SortedSideTable wrote the const workaround out by
hand: two public overloads delegating to a static helper templated on the
object, under a comment naming the trick. Every engine then repeated the pair
for each payload accessor. Both go, and a registry accessor is now

[[nodiscard]] auto &text_element_at(this auto &self, const ElementIdentifier id) {
  return self.m_texts.at(id);
}

pdf's Array and Dictionary lose eleven more. −238/+118 lines.
ooxml_text_list's Y-combinator stays, for ceiling 3 above, and says so.

resize_and_overwrite — 12.3 → 16.7 GB/s on byte_stream::read_u8s, the
chunked read every binary engine goes through, measured against an in-memory
stream (which is what a zip entry is). resize zero-fills a tail the very next
read overwrites. The callback must not throw, so a short read shrinks the string
back and the throw happens at the call site. xls_io's string body is left
alone — its read_bytes throws from inside, and a non-throwing path just for it
is not worth the buffer it saves.

ranges::to for the two loops that did nothing but map a range into a vector
fed straight into one call.

The public headers stay C++20

Nothing propagates the standard — no target_compile_features(odr PUBLIC …), no
cppstd in package_info — so a consumer keeps whatever standard it picked, as
long as src/odr/*.hpp stays C++20. Nothing checked that, and a this auto &self
in a public header would have broken someone else's build rather than ours, so
the last commit adds odr_public_headers_cpp20: all seventeen public headers in
one TU at cxx_std_20. One object file, no test data, no link, builds by
default. It was verified to fail on a deducing-this construct and pass again
without it.

conanfile.py's check_min_cppstd moves to 23 but sits in validate_build,
which constrains building odr rather than using it. That is the one
consumer-visible change, and it has a CHANGELOG entry.

Not done, on purpose

• std::format — ceiling 2, and off the table until the apple deployment
  targets reach macOS 13.3 / iOS 16.3. fix(html): keep the host's locale out of the css we write #828 already took the locale-safety it
  would have bought, through fmt. String building stays + / std::to_string;
  fmt is no faster at those lengths.
• std::unreachable — the switch defaults throw by design (AGENTS.md Fail
  fast); swapping a throw for UB is backwards.
• std::to_underlying — there is no static_cast<…>(std::underlying_type_t<…>)
  anywhere.
• std::byteswap — util::byte::{from,to}_{little,big}_endian is already
  byte-wise and portable, which is what keeps the big-endian gap tractable.
• Multidimensional operator[] — Sheet::cell(column, row) would read as
  sheet[column, row], but it is public API and a C++23-only spelling would push
  the standard onto consumers.